Development of a deep-learning model for classification of intensivists' extubation decisions using ventilator graphic monitor images: a prospective observational study.

Abstract

Background: Timely extubation requires advanced expertise and extensive clinical experience; however, intensive care specialists are scarce globally, and interpreting ventilator graphic monitor waveforms can be challenging for less experienced healthcare providers. Given these challenges, we aimed to develop a deep-learning model to classify extubation decisions by intensivists based on ventilator graphic monitor images obtained at the time of the spontaneous breathing trial (SBT) outcome determination.

Methods: This single-center prospective observational study was conducted in the Intensive Care Unit of Yamagata University Hospital between August 2023 and July 2025. Adult patients aged ≥ 20 years who received mechanical ventilation and underwent SAT and SBT using a Puritan Bennett^™ 980 ventilator were included. Ventilator graphic monitor images were acquired at the time of SBT outcome determination using the screen capture function. Four convolutional neural network architectures pretrained on ImageNet (DenseNet121, EfficientNet-B0, MobileNetV3-Large, and ResNet18) were evaluated using transfer learning. This study was divided into four periods, with a threefold cross-validation performed on the first three periods and the fourth period used as an independent test set. Gradient-weighted class activation mapping (Grad-CAM) was performed to visualize the basis for the model decisions.

Results: A total of 191 patients were included in the final analysis: 160 (83.8%) in the extubated group and 31 (16.2%) in the non-extubated group. MobileNetV3-Large demonstrated the highest performance, with an area under the curve (AUC) of 0.900 (95% confidence interval 0.807-0.993), sensitivity of 0.900, specificity of 0.727, and accuracy of 0.759. All four architectures exhibited consistently high performance (AUC: 0.827-0.900). The Grad-CAM analysis suggested that the model tended to focus on the entire waveform, including pressure, flow, and volume, in the extubated group while focusing primarily on the flow waveform in the non-extubated group.

Conclusions: This study developed a deep-learning model capable of classifying intensivists' extubation decisions from ventilator graphic monitor images with good classification performance (AUC 0.900). As a proof of concept, this approach suggests that waveform-based components of intensivists' decision-making may be approximated from image data, although further validation with larger multicenter datasets is required.

Trial registration: The study was registered in the UMIN Clinical Trials Registry (UMIN000051938).